ilessio-aiflowlab/project_slam_mipslam

SLAM-MIPSLAM — ANIMA Module

Part of the ANIMA Perception Suite by Robot Flow Labs.

Paper

MipSLAM: Alias-Free Gaussian Splatting SLAM (arXiv:2603.06989)

Yingzhao Li, Yan Li, Shixiong Tian, Yanjie Liu, Lijun Zhao, Gim Hee Lee

Architecture

MipSLAM is an RGB-D 3D Gaussian Splatting SLAM system with two key innovations:

1. Elliptical Adaptive Anti-aliasing (EAA) — alias-free rendering under changed camera configurations via importance-weighted elliptical quadrature (Eq. 5-6)
2. Spectral-Aware Pose Graph Optimization (SA-PGO) — trajectory drift reduction using frequency-domain analysis and graph Laplacian decomposition (Eq. 8-14)

The system tracks RGB-D frames, builds a dense Gaussian map, and supports re-rendering at arbitrary resolutions without aliasing artifacts.

Exported Formats

Format	File	Size	Use Case
PyTorch (.pth)	pytorch/slam_mipslam_v1.pth	~50 MB	Training, fine-tuning
SafeTensors	pytorch/slam_mipslam_v1.safetensors	12.9 MB	Fast loading, safe deserialization
ONNX	onnx/slam_mipslam_v1.onnx	2.3 KB	Cross-platform Gaussian projection
TensorRT FP16	tensorrt/slam_mipslam_v1_fp16.trt	—	Edge deployment (generate on target)
TensorRT FP32	tensorrt/slam_mipslam_v1_fp32.trt	—	Full precision (generate on target)

Evaluation Results (Replica Benchmark)

Scene	ATE RMSE (cm)	Gaussians	FPS
room0	128.28	200K	2.9
room1	157.64	200K	2.9
room2	133.26	200K	2.8
office0	252.36	200K	3.0
office1	217.85	200K	3.0
apartment_0	195.66	200K	2.9
Average	180.84	200K	2.9

Paper target: ATE RMSE 0.28 cm (with iterative pose refinement not yet implemented).

Usage

import torch
from safetensors.torch import load_file

# Load Gaussian map
data = load_file("pytorch/slam_mipslam_v1.safetensors")
means = data["gaussian_means"]      # [N, 3]
covs = data["gaussian_covs"]        # [N, 3, 3]
opacities = data["gaussian_opacities"]  # [N]
colors = data["gaussian_colors"]    # [N, 3]
trajectory = data["trajectory"]     # [T, 4, 4]

CUDA Kernels

Shared EAA CUDA kernel available at: /mnt/forge-data/shared_infra/cuda_extensions/eaa_renderer/

from eaa_renderer import eaa_render
rgb, depth, alpha = eaa_render(means2d, covs2d, depths, opacities, colors, H=480, W=640)

Training

• Hardware: NVIDIA L4 (23 GB VRAM)
• Framework: PyTorch 2.11 + custom EAA CUDA kernel
• Dataset: Replica RGB-D (6 sequences, 500 frames each, 480×640)
• Config: See configs/training.toml

Docker

docker compose -f docker-compose.serve.yml --profile serve up -d
curl http://localhost:8080/health

License

Apache 2.0 — Robot Flow Labs / AIFLOW LABS LIMITED